← All education signals
Governance

The Underage Dataset

Children’s and students’ data is being fenced off from AI training. The amended US privacy rule treats training as a use needing separate consent. California bars it outright.

What is it?

Data about children and students is being walled off from AI training. The federal children’s privacy rule makes training a use that needs its own parental consent. California goes further and bans edtech operators from training generative AI on student records.

Why now?

Compliance with the amended COPPA Rule began on 22 April 2026. On 10 September California signed AB 1159, which reaches any company that knows its product is used in schools. The ban arrives with a private right to sue attached.

Who’s building it?

California legislators, the Federal Trade Commission, edtech operators, general platforms used in classrooms, and parents.

California LegislatureFTCEdtech operatorsClassroom platformsTeachers' unions
The full read
Why now, who is building it, where the market goes, and what stays unsettled.
Get weekly →
01 · The market
Who’s in it
Builders — edtech operators and general platforms whose products reach schools.
Buyers — schools, pupils and parents whose records flow through classroom products every day.
Gatekeepers — California courts, the FTC and parents holding consent.
What’s at stake
Whether a child’s or student’s record can ever enter a training set once it reaches a school product.
Whether general platforms accept edtech rules in classrooms.
Whether a parent can still withdraw consent after training.
What’s unsettled
How consent applies to children’s data collected before the rule.
What a model must do when consent is withdrawn.
Whether any FTC action names model training as the use that broke the rule.
02 · How the demand matureda progression, not a timeline
Novelty
Training on children’s data is named as its own use needing consent.
Retention
California removes the consent route for student data and bans training outright.
Willingness to pay
Operators face $500 per violation or actual damages, whichever is greater, if a student sues.
Category
Age and school origin become attributes every training record must carry.
03 · Where the market goesthe forks, not a prediction
→The clean pool: operators strip child and student records from training at intake, and education AI learns from other material.
→Other states copy California’s ban, and the consent route narrows across the country.
→The first private suit tests what counts as developing an AI system under the law.
→A parent withdraws consent, and an already trained model must answer for it.
Everything hangs on whether a rule written for data at intake can reach a model that has already learned from that data.
04 · The Consented Data deskadjacent market, moving now
The Forgetting Right
Making trained models unlearn a person.
Deletion must reach machine unlearning, often infeasible for large systems.
California Lawyers Association
The Training Consent Wars
Who agrees before a model learns from them.
Anthropic’s $1.5B author settlement won final approval in 2026.
TechCrunch
Post-Feed Kids
A generation now growing up beyond the feed.
FTC opened a Section 6(b) inquiry into chatbots and children.
Tech Policy Press
05 · The Wire
01
AB 1159 bars training AI on student data.
California Legislature ↗
02
California bars tech firms from training AI on student data
CalMatters ↗
03
Compliance with the amended COPPA Rule is required from 22 April 2026, including separate parental consent for third-party disclosure.
Hunton ↗
04
The FTC’s commentary treats disclosures to train or develop AI as not integral to a service, so consent is required.
Data Protection Report ↗
05
California’s student AI-training ban starts 1 January 2027 for K-12
Venable LLP ↗
The panel on this signal
They agreed on
–California bans edtech operators from training generative AI on student data, with no consent route around it.
–The federal rule asks for separate parental consent.
–No FTC enforcement naming model training has been found, so the federal rule is untested.
The machine-unlearning researcher disagreed on
–A ban at intake cannot remove student data already inside trained models, and reliable unlearning at scale does not yet exist.
Know more about the method →
✦How this signal was built
01 · Panel
A Delphi panel of 4 AI analysts, each reasoning independently.
→
02 · Debate
They challenge and revise each other’s reads over two rounds.
→
03 · Evidence
Every claim is tied to a cited, linked source on The Wire.
→
04 · Synthesis
A moderator writes it into one read, disagreement and all.